school of electrical engineering telecommunications unsw cost-effective broadcast for fully...
DESCRIPTION
School of Electrical Engineering &Telecommunications UNSW Wireless Environments Characterized by –Highly transient node populations –Wide range of users form non cooperating organizations –Searches on partial information –Not typically looking for “rare” information – replicated at a number of places Not a good match for structured systems –Back to unstructured systemsTRANSCRIPT
School of Electrical Engineering &
Telecomm
unications
UNSW
Cost-effective Broadcast for Fully Decentralized Peer-to-peer Networks
Marius Portmann & Aruna Seneviratne
School of Electrical Engineering &
Telecomm
unications
UNSW
Peer to Peer Systems• Two types
– Structured•Guarantee location of content (if exists)•Access within bounded number of hopsControl of data placements and topology
– Unstructured•Decentalized•Looser guaranteesPlacement of data and topology is ad-hoc
School of Electrical Engineering &
Telecomm
unications
UNSW
Wireless Environments• Characterized by
– Highly transient node populations– Wide range of users form non cooperating
organizations– Searches on partial information– Not typically looking for “rare” information
– replicated at a number of places• Not a good match for structured
systems – Back to unstructured systems
School of Electrical Engineering &
Telecomm
unications
UNSW
Unstructured Systems• Most widely used application of
p2p systems is file sharing– As the placement of data is ad-hoc
• Only random searchers are possible• Hard to find desired files without wide
distribution of queries• Unscalable unless can improve the
efficiency of searches
School of Electrical Engineering &
Telecomm
unications
UNSW
Example - Gnutella• Gnutella can be considered as “pure”
peer-to-peer system – Fully decentralized and distributed searching
• Operation of Gnutella– Two types of services
• Searching for files • Peer discovery
– Implemented with application level broadcasts
– Broadcast is implemented with TTL flooding
School of Electrical Engineering &
Telecomm
unications
UNSW
File Location• A query message is forwarded to all its
neighbors, except for the one, where it was received from
• Each message has a Time To Live (TTL) – Decremented by one at each visited node – Message is dropped when TTL=0
• Each message has an unique ID • Node keeps a record of IDs of messages
that it has seen in the recent past – Message with the same ID and type as ones
that that have been received are dropped
School of Electrical Engineering &
Telecomm
unications
UNSW
Cost Metric 1• Define a cost metric for
comparison of methods– number of messages that are
generated and forwarded – based solely on the network size and
the average node degree, • Estimate the average bandwidth
consumption per node
School of Electrical Engineering &
Telecomm
unications
UNSW
Cost Metric 2
network in the nodes ofnumber :Ni nodeby forwarded messages ofnumber :
broadcast aby reached nodes ofnumber :
1 1
i
N
ii
mr
mr
c
School of Electrical Engineering &
Telecomm
unications
UNSW
Flooding - Unscalable• Resource consumption per node of
flooding based broadcast can be prohibitively high, even for networks of moderate size
School of Electrical Engineering &
Telecomm
unications
UNSW
Rumor Mongering or Gossip Protocols
• A class of probabilistic protocols for message routing
• Messages are spread in a network much like a disease in a susceptible population. (epidemiological protocol)
• The neighbors to which messages are forwarded to by each node are chosen randomly.
• Trades off reliability and speed for a reduction in cost
School of Electrical Engineering &
Telecomm
unications
UNSW
Blind Counter Rumor Mongering
• A node n initiates a broadcast – Send the message m to B neighbors,
chosen at random– When a node (p) receives a message m
from anther node (q)• If (p has received m no more than F times)• p sends m to B uniformly randomly chosen
neighbors that p knows have not yet seen m– p knows if its neighbor q has already seen the
message m only if p has sent it to q previously, or if p received the message from q
School of Electrical Engineering &
Telecomm
unications
UNSW
Cost of BCRM• Difficult to obtain analytical
expressions to describe the behavior of a Gossip protocol, even for relatively simple topologies
• Can give an upper limit – bounded by BF- an upper limit for
the cost c
School of Electrical Engineering &
Telecomm
unications
UNSW
Simulation Results• Barabási Topology:
– Model for generating topology is based on how typical p2p networks evolve
– Power-law characteristics• 1000 nodes with an average node
degree of 6 – F and B for the BCRM was set to be 2
School of Electrical Engineering &
Telecomm
unications
UNSW
Some More Results
• Trade-off of cost, reliability and time by choosing F and B appropriately
• Level of cost reduction depends on the average node degree– The higher the node degree is, the bigger
the potential for cost reduction
School of Electrical Engineering &
Telecomm
unications
UNSW
P2P Network Topologies• Typical characteristic of peer-to-peer networks
is a power-law distribution of the node degrees– most nodes have few links while a small number of
nodes have a large number of links
From Matei Ripeanu & Ian Foster
School of Electrical Engineering &
Telecomm
unications
UNSW
Deterministic Rumor Mongering
• Make intelligent decisions as to which of its neighbors to forward messages to
• Based it on the node degree of the corresponding nodes– The nodes with the lowest degree are
chosen first
School of Electrical Engineering &
Telecomm
unications
UNSW
Deterministic Rumor Mongering cont.
• When a node p receives a message m from node q
– If (p has received m no more than F times)
1) send m to all of its neighbors of degree one, and
2) B of the rest of its neighbors with the lowest node degree, that p knows have not yet seen m
School of Electrical Engineering &
Telecomm
unications
UNSW
Rationale for (1)• Pendant neighbors, have no other
chance to receive the message• These pendant neighbors cannot
contribute to the further propagation of the message – not considered for the limit of B
messages to be forwarded
School of Electrical Engineering &
Telecomm
unications
UNSW
Rationale for (2)• Nodes of high degree receive a large
number of copies of the same message– This overhead grows approximately linearly
with the node degree– Also with higher parameters B and F.
0
10
20
30
40
50
60
70
80
90
100
0 20 40 60 80 100 120
Node Degree
Mes
sage
s Re
ceiv
ed
B=3, F=2
B=2, F=2
B=2, F=1
School of Electrical Engineering &
Telecomm
unications
UNSW
Viability• The only requirement is that each node
knows the node degree of its immediate neighbors
• Not in conflict with the decentralized nature of the networks – Can easily be integrated– Gnutella
• a one byte field in the Gnutella message header • Increasing the minimal message size by less than
5%.
School of Electrical Engineering &
Telecomm
unications
UNSW
Some Results
Performance of Deterministic Rumor Mongering compared to Blind Counter Rumor Mongering
• For a given B and F, DRM achieves a significant higher reach than the BCRM, within a shorter time
• For a given reach, DRM has a significantly lower cost
B Freach time cost c reach time cost c
2 1 67.7% 23.2 2.00 96.10% 21.1 2.003 1 86.9% 17.1 2.65 99.00% 16.4 2.602 2 91.7% 21.6 2.78 99.80% 18.4 2.822 3 97.1% 19.9 3.17 100.00% 18.1 3.303 2 97.7% 15.3 3.41 100.00% 14.2 3.423 3 99.2% 14.4 3.73 100.00% 14.0 3.76
BCRM DRM
School of Electrical Engineering &
Telecomm
unications
UNSW
(3,2)(3,3)
(2,3)(2,2)(3,1)
(2,1)
(3,2)(2,3)
(3,3)
(2,2)(3,1)
(2,1)
00.5
11.5
22.5
33.5
44.5
5
60 70 80 90 100
Reach (%)
Cos
t per
Nod
e re
ache
dSome More Results
BCRM DRM
(B,F)
School of Electrical Engineering &
Telecomm
unications
UNSW
Conclusions• Unstructured peer-to-peer systems are more
suitable for wireless environments• For unstructured systems to be viable,
scalable methods of searching need to developed
• The obvious way of is to look at alternatives to broadcast
• One such scheme that have been used in the past in other application is Rumor Mongering (Gossiping)
• We show that, Rumor Mongering, can be used as a basis for providing an alternative flooding for distributing queries in unstructured peer to peer systems
School of Electrical Engineering &
Telecomm
unications
UNSW
More InformationAvailable form mobqos.ee.unsw.edu.au
M. Portmann, Pipat Sookavatna, Sebstien Ardon and Aruna Seneviratne,”The Cost of Peer Discovery and Searching in the Gnutella Peer-to-peer File Sharing Protocol”, IEEE ICON 2001, Bangkok, September 2001M. Portmann, and Aruna Seneviratne, “The Cost of Application-level Broadcast in a fully Decentralized Peer-to-peer Networks”, ISCC, Italy, July 2002M. Portmann, and Aruna Seneviratne, “Cost-effective Broadcast for Fully Decentralized Peer-to-peer Networks”, accepted for publication, Computer Communication, Special Issue on Ubiquitous Computing
Also related workQin Lv, Sylvia Ratnasamy and Scott Shenker,”Can Heterogeneity Make Gnutella Scalable?”, 1st International Workshop on Peer-to-Peer Systems (IPTPS '02), Cambridge, MA, USA, March 2002Berverly Yang, and Hector Garcia-Molina,”Efficient Search in Peer-to-Peer Networks”, 1st International Workshop on Peer-to-Peer Systems (IPTPS '02), Cambridge, MA, USA, March 2002
School of Electrical Engineering &
Telecomm
unications
UNSW
Possibly some …..
?